Memory that persists
across every session.
NockBrain is memory you approve for Claude Code. Decisions, directives, and corrections you review go into a signed store on your machine. When a prompt asks about the past, the opt-in recall hook hands the matching facts back to Claude. No network calls, no model calls, no telemetry.
Animated demo: an agent starts work and NockBrain recalls the accepted memories that match the prompt.
Three kinds
of memory.
NockBrain keeps a local store of plain JSON files. Every record is signed and stays on your machine: no cloud, no lock-in.
What you settled and why. Tag a line in your notes and it becomes a candidate fact that points back to the line it came from.
# notes.md - [DECISION] Invoices round per line, not per total.
Standing instructions you want kept, recalled when a later question about past work shares their words.
# notes.md - [DIRECTIVE] Run the tests before opening a PR.
Mistakes you caught once and never want repeated. Tag them like any other fact and apply them after review.
# notes.md - [CORRECTION] The staging database is read-only.
One minute,
start to finish.
What NockBrain does for you, how it works, and the one limit to know before you try it. Every claim in the narration points at the line of code behind it.
Watch the explainer on GitHub ↗ · 57 seconds, captions included
Setup is one script
and a local store.
A Python CLI that uses only the standard library, and a signed store on your machine. Two optional Claude Code hooks queue candidates for review and recall the facts you approved, and a read-only MCP server lets other agents recall from the same store.
Initialize the store
python3 bin/consumer-brain.py init --store "$HOME/nockbrain-store" creates a fresh store with its own signing key, in a folder only your user can open. No accounts, no cloud service.
Propose, review, apply
Notes you choose, or the optional hook that queues what was said in a session, become candidates in a review queue with secrets redacted. Nothing is in memory until you read the review and apply it.
Recall when a prompt asks about the past
Recall is opt-in. The hook checks every signature in the store first, then ranks your current facts by the words they share with your question (BM25) and hands Claude up to 800 tokens, labeled as reference.
Correct or forget on purpose
Replace a wrong fact or erase one record through the same review. A correction keeps the old record as history. Nothing is rewritten or deleted behind your back.
Recall from other agents
A read-only MCP server lets Codex CLI, Gemini CLI, Cursor and Claude Desktop look things up in the same store, with the same signature checks and 800-token budget. It has no write tools, and capture stays Claude Code only. Setup for each client.
Install
macOS or Linux with Python 3.9 or newer. NockBrain uses only the standard library, so there is nothing else to install. Clone it somewhere you'll keep it: the hook settings point at this folder.
# clone the repo $ git clone https://github.com/nocktechnologies/nockbrain.git $ cd nockbrain # create a store (the path must be new) $ STORE="$HOME/nockbrain-store" $ python3 bin/consumer-brain.py init --store "$STORE" # propose facts from notes, read the review, then apply it $ python3 bin/consumer-brain.py propose --store "$STORE" --format markdown --source "$PWD/examples/notes.md" $ python3 bin/consumer-brain.py review --store "$STORE" --proposal FULL_DIGEST $ python3 bin/consumer-brain.py apply --store "$STORE" --proposal FULL_DIGEST # browse your facts in the read-only Memory Explorer $ python3 bin/consumer-brain.py open --store "$STORE" # turn on capture and recall for one Claude Code project $ python3 bin/consumer-brain.py setup-hooks --store "$STORE" --transcript-root "$HOME/.claude/projects/YOUR-PROJECT-FOLDER" $ claude --settings "$STORE/claude-settings.json"
propose prints the full digest and the exact review command. Read the whole review before you apply it. Launch Claude Code without the settings file and NockBrain stays out of the session.
Context that outlasts
the session.
A first release for one person and one store. Recall matches words, so a fact phrased differently from your question won't come back.
Stop re-explaining
the project.
Give Claude a memory.
The rest of the fleet.
NockBrain is the memory. It sits alongside the fence, the playbook, and the dashboard that ties the fleet together.
Filesystem, network, and secret fencing for AI agents. Every limit stated per platform.
Learn more → FREE · OSS Nock Skills · the playbookOperational patterns: handoff protocols, review pipelines, standing orders.
Learn more → PRIVATE ALPHA Nock Terminal · the cockpitMission control for terminal-native coding agents (Claude Code, Codex CLI, Gemini CLI). Tabs, status, git, command palette. macOS, Windows, Linux. Private alpha.
Learn more → PRIVATE ALPHA · $49/MO AT LAUNCH Nock Command · the dashboardFleet coordination. Sprint planning, /fire dispatch, agent messaging, content pipeline.
Learn more → ← Back to homepage